hw/null-machine: Add the kvm_type() hook for MIPS
PATCH v2 1/1 hw/null-machine: Add the kvm_type() hook for MIPS
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the kvm_type() hook to the null-machine.
This seems not a very good solution, but I cannot do it better now.
--- hw/core/meson.build | 2 +- hw/core/null-machine.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)
hw/core/meson.build +1 -1
@@ -35,7 +35,6 @@ softmmu_ss.add(files( 'machine-hmp-cmds.c', 'machine.c', 'nmi.c', - 'null-machine.c', 'qdev-fw.c', 'qdev-properties-system.c', 'sysbus.c',
@@ -45,5 +44,6 @@ softmmu_ss.add(files( specific_ss.add(when: 'CONFIG_SOFTMMU', if_true: files( 'machine-qmp-cmds.c', + 'null-machine.c', 'numa.c', ))
hw/core/null-machine.c +6 -0
@@ -17,6 +17,9 @@#include "sysemu/sysemu.h" #include "exec/address-spaces.h" #include "hw/core/cpu.h" +#ifdef TARGET_MIPS +#include "kvm_mips.h" +#endif static void machine_none_init (MachineState *mch) {
@@ -55,6 +58,9 @@ static void machine_none_machine_init(MachineClass *mc) mc->no_floppy =1 ; mc->no_cdrom =1 ; mc->no_sdcard =1 ; +#ifdef TARGET_MIPS + mc->kvm_type = mips_kvm_type; +#endif }DEFINE_MACHINE ("none" , machine_none_machine_init)
Signature · 1 line
2.7.0
Hi Huacai,
Quoted 49 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
Reviewed-by Aleksandar Markovic <[email protected]>Signed-off-by Huacai Chen <[email protected]>Co-developed-by Jiaxun Yang <[email protected]>--- hw/core/meson.build | 2 +- hw/core/null-machine.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)hw/core/meson.build +1
-1@@ -35,7 +35,6 @@ softmmu_ss.add(files( 'machine-hmp-cmds.c', 'machine.c', 'nmi.c', - 'null-machine.c', 'qdev-fw.c', 'qdev-properties-system.c', 'sysbus.c',@@ -45,5 +44,6 @@ softmmu_ss.add(files( specific_ss.add(when: 'CONFIG_SOFTMMU', if_true: files( 'machine-qmp-cmds.c', + 'null-machine.c', 'numa.c', ))hw/core/null-machine.c +3
-0@@ -17,6 +17,9 @@#include "sysemu/sysemu.h" #include "exec/address-spaces.h" #include "hw/core/cpu.h" +#ifdef TARGET_MIPS +#include "kvm_mips.h" +#endif static void machine_none_init (MachineState *mch) {
hw/core/null-machine.c static void machine_none_machine_init(MachineClass *mc) · line 63
Quoted 6 lines
hw/core/null-machine.c +3
-0@@ -55,6 +58,9 @@ static void machine_none_machine_init(MachineClass *mc) mc->no_floppy =1 ; mc->no_cdrom =1 ; mc->no_sdcard =1 ; +#ifdef TARGET_MIPS + mc->kvm_type = mips_kvm_type; +#endif
I'm a bit confused here, you are taking the same path from your v4...
https://www.mail-archive.com/[email protected]/msg712550.html
Did you rebase the correct version?
Quoted 4 lines
hw/core/null-machine.c +0
-0}DEFINE_MACHINE ("none" , machine_none_machine_init)
Hi, Philippe,
Quoted 63 lines · Philippe Mathieu-Daudé
Hi Huacai,
Quoted 49 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
Reviewed-by Aleksandar Markovic <[email protected]>Signed-off-by Huacai Chen <[email protected]>Co-developed-by Jiaxun Yang <[email protected]>--- hw/core/meson.build | 2 +- hw/core/null-machine.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)hw/core/meson.build +1
-1@@ -35,7 +35,6 @@ softmmu_ss.add(files( 'machine-hmp-cmds.c', 'machine.c', 'nmi.c', - 'null-machine.c', 'qdev-fw.c', 'qdev-properties-system.c', 'sysbus.c',@@ -45,5 +44,6 @@ softmmu_ss.add(files( specific_ss.add(when: 'CONFIG_SOFTMMU', if_true: files( 'machine-qmp-cmds.c', + 'null-machine.c', 'numa.c', ))hw/core/null-machine.c +3
-0@@ -17,6 +17,9 @@#include "sysemu/sysemu.h" #include "exec/address-spaces.h" #include "hw/core/cpu.h" +#ifdef TARGET_MIPS +#include "kvm_mips.h" +#endif static void machine_none_init (MachineState *mch) {hw/core/null-machine.c static void machine_none_machine_init(MachineClass *mc) · line 63
Quoted 6 lines
hw/core/null-machine.c +3
-0@@ -55,6 +58,9 @@ static void machine_none_machine_init(MachineClass *mc) mc->no_floppy =1 ; mc->no_cdrom =1 ; mc->no_sdcard =1 ; +#ifdef TARGET_MIPS + mc->kvm_type = mips_kvm_type; +#endif I'm a bit confused here, you are taking the same path from your v4...
https://www.mail-archive.com/[email protected]/msg712550.html
Quoted 1 line
Did you rebase the correct version?
The old patch has split to two parts, the first part is used by MIPS machine and already merged. This is the second part used by the null-machine (and libvirt use null-machine to detect kvm capabilities).
Huacai
Quoted 5 lines
Quoted 5 lines
hw/core/null-machine.c +0
-0}DEFINE_MACHINE ("none" , machine_none_machine_init)
You forgot to Cc the maintainers, doing it for you:
./scripts/get_maintainer.pl -f hw/core/null-machine.c Eduardo Habkost <[email protected]> (supporter:Machine core) Marcel Apfelbaum <[email protected]> (supporter:Machine core)
Quoted 79 lines · Huacai Chen
Hi, Philippe,
Quoted 63 lines · Philippe Mathieu-Daudé
Hi Huacai,
Quoted 49 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
Reviewed-by Aleksandar Markovic <[email protected]>Signed-off-by Huacai Chen <[email protected]>Co-developed-by Jiaxun Yang <[email protected]>--- hw/core/meson.build | 2 +- hw/core/null-machine.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)hw/core/meson.build +1
-1@@ -35,7 +35,6 @@ softmmu_ss.add(files( 'machine-hmp-cmds.c', 'machine.c', 'nmi.c', - 'null-machine.c', 'qdev-fw.c', 'qdev-properties-system.c', 'sysbus.c',@@ -45,5 +44,6 @@ softmmu_ss.add(files( specific_ss.add(when: 'CONFIG_SOFTMMU', if_true: files( 'machine-qmp-cmds.c', + 'null-machine.c', 'numa.c', ))hw/core/null-machine.c +3
-0@@ -17,6 +17,9 @@#include "sysemu/sysemu.h" #include "exec/address-spaces.h" #include "hw/core/cpu.h" +#ifdef TARGET_MIPS +#include "kvm_mips.h" +#endif static void machine_none_init (MachineState *mch) {hw/core/null-machine.c static void machine_none_machine_init(MachineClass *mc) · line 63
Quoted 6 lines
hw/core/null-machine.c +3
-0@@ -55,6 +58,9 @@ static void machine_none_machine_init(MachineClass *mc) mc->no_floppy =1 ; mc->no_cdrom =1 ; mc->no_sdcard =1 ; +#ifdef TARGET_MIPS + mc->kvm_type = mips_kvm_type; +#endif I'm a bit confused here, you are taking the same path from your v4...
https://www.mail-archive.com/[email protected]/msg712550.htmlQuoted 1 line
Did you rebase the correct version?
The old patch has split to two parts, the first part is used by MIPS machine and already merged. This is the second part used by the null-machine (and libvirt use null-machine to detect kvm capabilities).
Huacai
Quoted 5 lines
Quoted 5 lines
hw/core/null-machine.c +0
-0}DEFINE_MACHINE ("none" , machine_none_machine_init)
Hi, Philippe,
Quoted 6 lines · Philippe Mathieu-Daudé
You forgot to Cc the maintainers, doing it for you:
./scripts/get_maintainer.pl-fhw/core/null-machine.cEduardo Habkost <[email protected]> (supporter:Machine core) Marcel Apfelbaum <[email protected]> (supporter:Machine core)
Thank you very much!
Huacai
Quoted 81 lines · Huacai Chen
Quoted 79 lines
Hi, Philippe,
Quoted 63 lines · Philippe Mathieu-Daudé
Hi Huacai,
Quoted 49 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
Reviewed-by Aleksandar Markovic <[email protected]>Signed-off-by Huacai Chen <[email protected]>Co-developed-by Jiaxun Yang <[email protected]>--- hw/core/meson.build | 2 +- hw/core/null-machine.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)hw/core/meson.build +1
-1@@ -35,7 +35,6 @@ softmmu_ss.add(files( 'machine-hmp-cmds.c', 'machine.c', 'nmi.c', - 'null-machine.c', 'qdev-fw.c', 'qdev-properties-system.c', 'sysbus.c',@@ -45,5 +44,6 @@ softmmu_ss.add(files( specific_ss.add(when: 'CONFIG_SOFTMMU', if_true: files( 'machine-qmp-cmds.c', + 'null-machine.c', 'numa.c', ))hw/core/null-machine.c +3
-0@@ -17,6 +17,9 @@#include "sysemu/sysemu.h" #include "exec/address-spaces.h" #include "hw/core/cpu.h" +#ifdef TARGET_MIPS +#include "kvm_mips.h" +#endif static void machine_none_init (MachineState *mch) {hw/core/null-machine.c static void machine_none_machine_init(MachineClass *mc) · line 63
Quoted 6 lines
hw/core/null-machine.c +3
-0@@ -55,6 +58,9 @@ static void machine_none_machine_init(MachineClass *mc) mc->no_floppy =1 ; mc->no_cdrom =1 ; mc->no_sdcard =1 ; +#ifdef TARGET_MIPS + mc->kvm_type = mips_kvm_type; +#endif I'm a bit confused here, you are taking the same path from your v4...
https://www.mail-archive.com/[email protected]/msg712550.htmlQuoted 1 line
Did you rebase the correct version?
The old patch has split to two parts, the first part is used by MIPS machine and already merged. This is the second part used by the null-machine (and libvirt use null-machine to detect kvm capabilities).
Huacai
Quoted 5 lines
Quoted 5 lines
hw/core/null-machine.c +0
-0}DEFINE_MACHINE ("none" , machine_none_machine_init)
Quoted 6 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
This is still ugly. Why do the other architectures do not have the same problem? Let's see... in kvm-all.c, we have:
int type = 0;
[...]
kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
if (mc->kvm_type) {
type = mc->kvm_type(ms, kvm_type);
} else if (kvm_type) {
ret = -EINVAL;
fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
goto err;
}
do {
ret = kvm_ioctl(s, KVM_CREATE_VM, type);
} while (ret == -EINTR);Thus the KVM_CREATE_VM ioctl is likely called with type = 0 in this case (i.e. when libvirt probes with the "null"-machine).
Now let's have a look at the kernel. The "type" parameter is passed there to the architecture specific function kvm_arch_init_vm(). For powerpc, this looks like:
if (type == 0) {
if (kvmppc_hv_ops)
kvm_ops = kvmppc_hv_ops;
else
kvm_ops = kvmppc_pr_ops;
if (!kvm_ops)
goto err_out;
} else if (type == KVM_VM_PPC_HV) {
if (!kvmppc_hv_ops)
goto err_out;
kvm_ops = kvmppc_hv_ops;
} else if (type == KVM_VM_PPC_PR) {
if (!kvmppc_pr_ops)
goto err_out;
kvm_ops = kvmppc_pr_ops;
} else
goto err_out;That means for type == 0, it automatically detects the best kvm-type.
For mips, this function looks like this:
switch (type) {#ifdef CONFIG_KVM_MIPS_VZ
case KVM_VM_MIPS_VZ:
#else
case KVM_VM_MIPS_TE:
#endif break; default: /* Unsupported KVM type */ return -EINVAL; };
That means, for type == 0, it returns -EINVAL here!
Looking at the API docu in Documentation/virt/kvm/api.rst the description of the type parameter is quite sparse, but it says:
"You probably want to use 0 as machine type."
So I think this is a bug in the implementation of KVM in the mips kernel code. The kvm_arch_init_vm() in the mips code should do the same as on powerpc, and use the best available KVM type there instead of returning EINVAL. Once that is fixed there, you don't need this patch here for QEMU anymore.
HTH,
Thomas
Quoted 52 lines
Reviewed-by Aleksandar Markovic <[email protected]>Signed-off-by Huacai Chen <[email protected]>Co-developed-by Jiaxun Yang <[email protected]>--- hw/core/meson.build | 2 +- hw/core/null-machine.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)hw/core/meson.build +1
-1@@ -35,7 +35,6 @@ softmmu_ss.add(files( 'machine-hmp-cmds.c', 'machine.c', 'nmi.c', - 'null-machine.c', 'qdev-fw.c', 'qdev-properties-system.c', 'sysbus.c',@@ -45,5 +44,6 @@ softmmu_ss.add(files( specific_ss.add(when: 'CONFIG_SOFTMMU', if_true: files( 'machine-qmp-cmds.c', + 'null-machine.c', 'numa.c', ))hw/core/null-machine.c +6
-0@@ -17,6 +17,9 @@#include "sysemu/sysemu.h" #include "exec/address-spaces.h" #include "hw/core/cpu.h" +#ifdef TARGET_MIPS +#include "kvm_mips.h" +#endif static void machine_none_init (MachineState *mch) {@@ -55,6 +58,9 @@ static void machine_none_machine_init(MachineClass *mc) mc->no_floppy =1 ; mc->no_cdrom =1 ; mc->no_sdcard =1 ; +#ifdef TARGET_MIPS + mc->kvm_type = mips_kvm_type; +#endif }DEFINE_MACHINE ("none" , machine_none_machine_init)
Quoted 74 lines · Thomas Huth
Quoted 6 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
This is still ugly. Why do the other architectures do not have the same problem? Let's see... in
kvm-all.c, we have:int type = 0; [...] kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type"); if (mc->kvm_type) { type = mc->kvm_type(ms, kvm_type); } else if (kvm_type) { ret = -EINVAL; fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type); goto err; } do { ret = kvm_ioctl(s, KVM_CREATE_VM, type); } while (ret == -EINTR);Thus the
KVM_CREATE_VMioctl is likely called with type = 0 in this case (i.e. when libvirt probes with the "null"-machine).Now let's have a look at the kernel. The "type" parameter is passed there to the architecture specific function
kvm_arch_init_vm(). For powerpc, this looks like:if (type == 0) { if (kvmppc_hv_ops) kvm_ops = kvmppc_hv_ops; else kvm_ops = kvmppc_pr_ops; if (!kvm_ops) goto err_out; } else if (type == KVM_VM_PPC_HV) { if (!kvmppc_hv_ops) goto err_out; kvm_ops = kvmppc_hv_ops; } else if (type == KVM_VM_PPC_PR) { if (!kvmppc_pr_ops) goto err_out; kvm_ops = kvmppc_pr_ops; } else goto err_out;That means for type == 0, it automatically detects the best kvm-type.
For mips, this function looks like this:
switch (type) {#ifdef
CONFIG_KVM_MIPS_VZ
caseKVM_VM_MIPS_VZ:
#else
caseKVM_VM_MIPS_TE:#endif break; default: /* Unsupported KVM type */ return -EINVAL; };That means, for type == 0, it returns
-EINVALhere!Looking at the API docu in
Documentation/virt/kvm/api.rstthe description of the type parameter is quite sparse, but it says:
Quoted 7 lines
"You probably want to use 0 as machine type."
So I think this is a bug in the implementation of KVM in the mips kernel code. The
kvm_arch_init_vm()in the mips code should do the same as on powerpc, and use the best available KVM type there instead of returning EINVAL. Once that is fixed there, you don't need this patch here for QEMU anymore.
If there's a way to make it work with older kernels, I assume we would still want to do it.
However, this kind of kvm-specific + arch-specific knowledge doesn't belong to machine core code. If we are going to add a #ifdef TARGET_MIPS to the code, we can simply do it inside kvm-all.c:kvm_init().
Signature · 1 line
Eduardo
Quoted 77 lines · Eduardo Habkost
Quoted 76 lines · Thomas Huth
Quoted 6 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
This is still ugly. Why do the other architectures do not have the same problem? Let's see... in
kvm-all.c, we have:int type = 0; [...] kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type"); if (mc->kvm_type) { type = mc->kvm_type(ms, kvm_type); } else if (kvm_type) { ret = -EINVAL; fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type); goto err; } do { ret = kvm_ioctl(s, KVM_CREATE_VM, type); } while (ret == -EINTR);Thus the
KVM_CREATE_VMioctl is likely called with type = 0 in this case (i.e. when libvirt probes with the "null"-machine).Now let's have a look at the kernel. The "type" parameter is passed there to the architecture specific function
kvm_arch_init_vm(). For powerpc, this looks like:if (type == 0) { if (kvmppc_hv_ops) kvm_ops = kvmppc_hv_ops; else kvm_ops = kvmppc_pr_ops; if (!kvm_ops) goto err_out; } else if (type == KVM_VM_PPC_HV) { if (!kvmppc_hv_ops) goto err_out; kvm_ops = kvmppc_hv_ops; } else if (type == KVM_VM_PPC_PR) { if (!kvmppc_pr_ops) goto err_out; kvm_ops = kvmppc_pr_ops; } else goto err_out;That means for type == 0, it automatically detects the best kvm-type.
For mips, this function looks like this:
switch (type) {#ifdef
CONFIG_KVM_MIPS_VZ
caseKVM_VM_MIPS_VZ:
#else
caseKVM_VM_MIPS_TE:#endif break; default: /* Unsupported KVM type */ return -EINVAL; };That means, for type == 0, it returns
-EINVALhere!Looking at the API docu in
Documentation/virt/kvm/api.rstthe description of the type parameter is quite sparse, but it says:"You probably want to use 0 as machine type."
Quoted 8 lines
Quoted 5 lines
So I think this is a bug in the implementation of KVM in the mips kernel code. The
kvm_arch_init_vm()in the mips code should do the same as on powerpc, and use the best available KVM type there instead of returning EINVAL. Once that is fixed there, you don't need this patch here for QEMU anymore.If there's a way to make it work with older kernels, I assume we would still want to do it.
IIUC this is available since kernel v5.8-rc1 (merged in 52cd0d972fa6) less than 3 months ago. v5.9-rc4 just got released, not sure if this can be fixed for the v5.9 release.
Is this small older kernels window important?
Huacai, can you amend in the commit description since when the feature is available in the Linux kernel please? (just for informative purpose).
Quoted 6 lines
However, this kind of kvm-specific + arch-specific knowledge doesn't belong to machine core code. If we are going to add a #ifdef
TARGET_MIPSto the code, we can simply do it inside kvm-all.c:kvm_init().
Hi, all,
Quoted 75 lines · Thomas Huth
Quoted 6 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
This is still ugly. Why do the other architectures do not have the same problem? Let's see... in
kvm-all.c, we have:int type = 0; [...] kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type"); if (mc->kvm_type) { type = mc->kvm_type(ms, kvm_type); } else if (kvm_type) { ret = -EINVAL; fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type); goto err; } do { ret = kvm_ioctl(s, KVM_CREATE_VM, type); } while (ret == -EINTR);Thus the
KVM_CREATE_VMioctl is likely called with type = 0 in this case (i.e. when libvirt probes with the "null"-machine).Now let's have a look at the kernel. The "type" parameter is passed there to the architecture specific function
kvm_arch_init_vm(). For powerpc, this looks like:if (type == 0) { if (kvmppc_hv_ops) kvm_ops = kvmppc_hv_ops; else kvm_ops = kvmppc_pr_ops; if (!kvm_ops) goto err_out; } else if (type == KVM_VM_PPC_HV) { if (!kvmppc_hv_ops) goto err_out; kvm_ops = kvmppc_hv_ops; } else if (type == KVM_VM_PPC_PR) { if (!kvmppc_pr_ops) goto err_out; kvm_ops = kvmppc_pr_ops; } else goto err_out;That means for type == 0, it automatically detects the best kvm-type.
For mips, this function looks like this:
switch (type) {#ifdef
CONFIG_KVM_MIPS_VZ
caseKVM_VM_MIPS_VZ:
#else
caseKVM_VM_MIPS_TE:#endif break; default: /* Unsupported KVM type */ return -EINVAL; };That means, for type == 0, it returns
-EINVALhere!Looking at the API docu in
Documentation/virt/kvm/api.rstthe description of the type parameter is quite sparse, but it says:
Quoted 7 lines
"You probably want to use 0 as machine type."
So I think this is a bug in the implementation of KVM in the mips kernel code. The
kvm_arch_init_vm()in the mips code should do the same as on powerpc, and use the best available KVM type there instead of returning EINVAL. Once that is fixed there, you don't need this patch here for QEMU anymore.
Yes, PPC use a good method, because it can use 0 as "automatic" #define KVM_VM_PPC_HV 1
#define KVM_VM_PPC_PR 2
Unfortunately, MIPS cannot do like this because it define 0 as "TE":
#define KVM_VM_MIPS_TE 0
#define KVM_VM_MIPS_VZ 1
So, it cannot be solved in kernel side, unless changing the definition of TE/VZ, but I think changing their definition is also unacceptable.
Huacai
Quoted 58 lines
HTH,
ThomasQuoted 52 lines
Reviewed-by Aleksandar Markovic <[email protected]>Signed-off-by Huacai Chen <[email protected]>Co-developed-by Jiaxun Yang <[email protected]>--- hw/core/meson.build | 2 +- hw/core/null-machine.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)hw/core/meson.build +1
-1@@ -35,7 +35,6 @@ softmmu_ss.add(files( 'machine-hmp-cmds.c', 'machine.c', 'nmi.c', - 'null-machine.c', 'qdev-fw.c', 'qdev-properties-system.c', 'sysbus.c',@@ -45,5 +44,6 @@ softmmu_ss.add(files( specific_ss.add(when: 'CONFIG_SOFTMMU', if_true: files( 'machine-qmp-cmds.c', + 'null-machine.c', 'numa.c', ))hw/core/null-machine.c +6
-0@@ -17,6 +17,9 @@#include "sysemu/sysemu.h" #include "exec/address-spaces.h" #include "hw/core/cpu.h" +#ifdef TARGET_MIPS +#include "kvm_mips.h" +#endif static void machine_none_init (MachineState *mch) {@@ -55,6 +58,9 @@ static void machine_none_machine_init(MachineClass *mc) mc->no_floppy =1 ; mc->no_cdrom =1 ; mc->no_sdcard =1 ; +#ifdef TARGET_MIPS + mc->kvm_type = mips_kvm_type; +#endif }DEFINE_MACHINE ("none" , machine_none_machine_init)
Signature · 1 line
Huacai Chen
Quoted 89 lines · chen huacai
Hi, all,
Quoted 75 lines · Thomas Huth
Quoted 6 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
This is still ugly. Why do the other architectures do not have the same problem? Let's see... in
kvm-all.c, we have:int type = 0; [...] kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type"); if (mc->kvm_type) { type = mc->kvm_type(ms, kvm_type); } else if (kvm_type) { ret = -EINVAL; fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type); goto err; } do { ret = kvm_ioctl(s, KVM_CREATE_VM, type); } while (ret == -EINTR);Thus the
KVM_CREATE_VMioctl is likely called with type = 0 in this case (i.e. when libvirt probes with the "null"-machine).Now let's have a look at the kernel. The "type" parameter is passed there to the architecture specific function
kvm_arch_init_vm(). For powerpc, this looks like:if (type == 0) { if (kvmppc_hv_ops) kvm_ops = kvmppc_hv_ops; else kvm_ops = kvmppc_pr_ops; if (!kvm_ops) goto err_out; } else if (type == KVM_VM_PPC_HV) { if (!kvmppc_hv_ops) goto err_out; kvm_ops = kvmppc_hv_ops; } else if (type == KVM_VM_PPC_PR) { if (!kvmppc_pr_ops) goto err_out; kvm_ops = kvmppc_pr_ops; } else goto err_out;That means for type == 0, it automatically detects the best kvm-type.
For mips, this function looks like this:
switch (type) {#ifdef
CONFIG_KVM_MIPS_VZ
caseKVM_VM_MIPS_VZ:
#else
caseKVM_VM_MIPS_TE:#endif break; default: /* Unsupported KVM type */ return -EINVAL; };That means, for type == 0, it returns
-EINVALhere!Looking at the API docu in
Documentation/virt/kvm/api.rstthe description of the type parameter is quite sparse, but it says:Quoted 7 lines
"You probably want to use 0 as machine type."
So I think this is a bug in the implementation of KVM in the mips kernel code. The
kvm_arch_init_vm()in the mips code should do the same as on powerpc, and use the best available KVM type there instead of returning EINVAL. Once that is fixed there, you don't need this patch here for QEMU anymore.Yes, PPC use a good method, because it can use 0 as "automatic" #define
KVM_VM_PPC_HV1
#defineKVM_VM_PPC_PR2
Quoted 6 lines
Unfortunately, MIPS cannot do like this because it define 0 as "TE":
#defineKVM_VM_MIPS_TE0
#defineKVM_VM_MIPS_VZ1So, it cannot be solved in kernel side, unless changing the definition of TE/VZ, but I think changing their definition is also unacceptable.
Ouch, ok, now I understood the problem. That sounds like a really bad decision on the kernel side.
But I think you could at least try to get it fixed on the kernel side:
Propose a patch to rename KVM_VM_MIPS_TE to KVM_VM_MIPS_DEFAULT and use 2 as new value for TE. The code that handles the default 0 case should then prefer TE over VZ, so that old userspace applications that provide 0 will continue to work as they did before, so I hope that the change is acceptable by the kernel folks. You should add a remark to the patch description that 0 is the established default value for probing in the QEMU/libvirt stack and that your patch is required on the kernel side to avoid ugly hacks in QEMU userspace code.
If they still reject your patch, I guess we have to bite the bullet and add your patch here...
Thanks,
Thomas
Hi, Thomas,
Quoted 111 lines · Thomas Huth
Quoted 89 lines · chen huacai
Hi, all,
Quoted 75 lines · Thomas Huth
Quoted 6 lines · Huacai Chen
MIPS has two types of KVM: TE & VZ, and TE is the default type. Now, libvirt uses a null-machine to detect the kvm capability. In the MIPS case, it will return "KVM not supported" on a VZ platform by default. So, add the
kvm_type()hook to the null-machine.This seems not a very good solution, but I cannot do it better now.
This is still ugly. Why do the other architectures do not have the same problem? Let's see... in
kvm-all.c, we have:int type = 0; [...] kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type"); if (mc->kvm_type) { type = mc->kvm_type(ms, kvm_type); } else if (kvm_type) { ret = -EINVAL; fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type); goto err; } do { ret = kvm_ioctl(s, KVM_CREATE_VM, type); } while (ret == -EINTR);Thus the
KVM_CREATE_VMioctl is likely called with type = 0 in this case (i.e. when libvirt probes with the "null"-machine).Now let's have a look at the kernel. The "type" parameter is passed there to the architecture specific function
kvm_arch_init_vm(). For powerpc, this looks like:if (type == 0) { if (kvmppc_hv_ops) kvm_ops = kvmppc_hv_ops; else kvm_ops = kvmppc_pr_ops; if (!kvm_ops) goto err_out; } else if (type == KVM_VM_PPC_HV) { if (!kvmppc_hv_ops) goto err_out; kvm_ops = kvmppc_hv_ops; } else if (type == KVM_VM_PPC_PR) { if (!kvmppc_pr_ops) goto err_out; kvm_ops = kvmppc_pr_ops; } else goto err_out;That means for type == 0, it automatically detects the best kvm-type.
For mips, this function looks like this:
switch (type) {#ifdef
CONFIG_KVM_MIPS_VZ
caseKVM_VM_MIPS_VZ:
#else
caseKVM_VM_MIPS_TE:#endif break; default: /* Unsupported KVM type */ return -EINVAL; };That means, for type == 0, it returns
-EINVALhere!Looking at the API docu in
Documentation/virt/kvm/api.rstthe description of the type parameter is quite sparse, but it says:Quoted 7 lines
"You probably want to use 0 as machine type."
So I think this is a bug in the implementation of KVM in the mips kernel code. The
kvm_arch_init_vm()in the mips code should do the same as on powerpc, and use the best available KVM type there instead of returning EINVAL. Once that is fixed there, you don't need this patch here for QEMU anymore.Yes, PPC use a good method, because it can use 0 as "automatic" #define
KVM_VM_PPC_HV1
#defineKVM_VM_PPC_PR2Quoted 6 lines
Unfortunately, MIPS cannot do like this because it define 0 as "TE":
#defineKVM_VM_MIPS_TE0
#defineKVM_VM_MIPS_VZ1So, it cannot be solved in kernel side, unless changing the definition of TE/VZ, but I think changing their definition is also unacceptable.
Ouch, ok, now I understood the problem. That sounds like a really bad decision on the kernel side.
But I think you could at least try to get it fixed on the kernel side:
Propose a patch to renameKVM_VM_MIPS_TEtoKVM_VM_MIPS_DEFAULTand use 2 as new value for TE. The code that handles the default 0 case should then prefer TE over VZ, so that old userspace applications that provide 0 will continue to work as they did before, so I hope that the change is acceptable by the kernel folks. You should add a remark to the patch description that 0 is the established default value for probing in the QEMU/libvirt stack and that your patch is required on the kernel side to avoid ugly hacks in QEMU userspace code.
Quoted 2 lines
If they still reject your patch, I guess we have to bite the bullet and add your patch here...
OK, let me try.
Huacai
Quoted 4 lines
Thanks,
Thomas